EXTREME POINTS OF VECTOR FUNCTIONS

SAMUEL KARLIN Several previous investigations have appeared in the literature which discuss the nature of the set T in »-dimensional space spanned by the vectors ifsdui, ■ ■ ■ , fsdp.n) where 5 ranges over all measur- able sets. It was first shown by Liapounov that xi ui, • • ■ , un are atomless measures, then the set F is convex and closed. Extensions of this result were achieved by D. Blackwell [l] and Dvoretsky, Wald, and Wolfowitz [2]. A study of the extreme points in certain function spaces is made from which the theorems concerning the range of finite vector measures are deduced as special cases. However, the extreme point theorems developed here have independent interest. In addition, some results dealing with infinite direct products of func- tion spaces are presented. Many of these ideas have statistical in- terpretation and applications in terms of replacing randomized tests by nonrandomized tests. 1. Preliminaries. Let p. denote a finite measure defined on a Borel field of sets $ given on an abstract set X. Let Liu, 5) denote the space of all integrable functions with respect to u. Finally, let Miß, %) be the space of all essentially bounded measurable functions defined on X. It is well known that Af(/i, 5) constitutes the conjugate space of Liu, %) and consequently the unit sphere of Af(/x, %) is bi- compact in the weak * topology [3]. Let (®Ln) denote the direct product of L taken with itself » times and take i®Mn) as the direct product of M n times. With appropriate choice of , i®M") becomes the conjugate space to (®7,n). Another description of i®Mn) is that it consists of all «-vectors, each component of which is an essentially bounded measurable function. In notation, let xEi®Mn) be denoted by x=(x,(/)), i=l, ■ • -, », with x,- in M. 2. Extreme points in Ma. Let A be a bounded closed in Euclidean »-space. Let Ma consist of all x in ( <8>Mn) whose range of values lie almost everywhere in A. I.e., for almost every t, x(i) = (x,(<)) is in A. Let B consist of the extreme points of A, and let B = closure of B. For n^3, it is not necessarily true that B = B. In view of the convexity of A, the set Ma is easily verified to be convex, bounded, and weak * closed. We indicate the proof of this last fact. Let x0 represent a weak * limit point of MA. Suppose xoEMa- This implies the existence of a set of positive finite measure Received by the editors October 11, 1952 and, in revised form, October 30, 1952. 603

License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use 604 SAMUEL KARLIN [August

E with x0(E)(¡.A. Since a denumerable number of hyperplanes de- termine A, it is easy to construct a plane {£<} so that Zt-i ¡tiVi>c lor {r]i} Ç_A while 23"_i £¿t?(Ei) ).

We note that (w-x)= ^2?_i JwiXidp represents the inner product and is èc for x in MA, while (wxa) 0 and suppose *oG-^B(í„> for every n. That is, except for a set En of p measure zero the range of values of x0 is in Mb^„). Since p( Z^") =0, we deduce that *0 is in Mr\B<.t^= Mb, which serves as a contradiction. Thus there exists a set E of positive meas- ure such that xo(t), lor t in E, is not in B(eo). It follows easily that a constant vector ä exists of small magnitude such that xo + â is in A when t is in E0(ZE. Put (â, l G Eo, $(t) = \ lo, t(£Eo. Consequently, Xo= (xo+)/2-\-(x0 —$)/2 with x0 + #, x0 —$ in MA. This impossibility establishes the result, Q.E.D. Remark. The set Mb need not be contained in (MA)U) (extreme points) but Mb is. 3. Extreme points of Ma with side conditions. A finite measure p is said to be atomless if for any measurable set E of positive meas- ure p there exists a measurable subset Eo such that 0

License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use 19531 EXTREME POINTS OF VECTORFUNCTIONS 605

less finite measures defined on the same Borel field of sets, then the span in Euclidean w space of the « tuples (/x(<)d/iy(i)) obtained by al- lowing x(£) to range over all characteristic functions is convex and closed. Therefore, it follows that the set of all » tuples ifxit)dp.jit)) with 0 =x(¿) ^ 1 spans the same set. An apparent weaker formulation of the above spanning result which is, however, equivalent is that for any given measurable set S there exists a set EES such that simultaneously UiiE)=miS)/2 for i=\, ■ ■ • , ». Let M [A, Uj] = [I,-,i= 1, • • • , »] with í,=/x(/)¿/tj(<) and x in Ma. It is of course understood that x is measurable with respect to all ßj. This set can also be viewed as points in Euclidean nm space. Theorem 2. 7/ju, (t'=l, ■ • • , ») consist of atomless finite measures, then the extreme points of the set T of all x(<) measurable /x,-(i = 1, • • •, w) satisfying xEMa and bi^fxdßi^äi are contained in Mb. Remark 1. The vector inequality fxdßi^äi means that the in- equality holds for each component of the vectors. Remark 2. The interest of Theorem 2 is that if we impose linear side conditions on the set Ma of the form 5¿^/xá/i¿^a¿, then no new extreme points are added to this convex subset of Ma when B = B. Only the set of extreme points of M a may be diminished. This result is in sharp contrast to the situation in the case of atomic meas- ures. We leave it to the reader to construct examples involving atomic measures which violate the conclusions of the theorem. Remark 3. If ß=ßX-\- ■ ■ • +/i„, then by virtue of the Radon- Nikodym theorem ßiiE) =fEfiit)dßit). The space Miß, %) considered here is the set of all bounded measurable functions x(¿), which is the conjugate space to 7,(/u, g) where % is the common Borel field of sets over which all the p¡ are defined. Therefore, bounded weak * closed sets in Miß, %) are bicompact and the same holds for (®M(ju, %)n). The statement of the theorem in terms of the /< is that the set Y of extreme points of all x(/) in MA which satisfy 5,¿/x/í(/)¿m(0 ^<*¿ is contained in Mb when ß is atomless. Since A is convex and closed and the linear restrictions are generated by elements of L iß, %), we deduce that T is bicompact and convex and hence possesses extreme points. The theorem thus determines the form of such extreme points. Proof of Theorem 2. Let x0 be an extreme point of T and suppose x0 is not in Mb. Then an argument as in Theorem 1 shows that for some positive eo the point x0 is not in ikfs(eo). Consequently there exists a set 5 of positive measure ß for which x0(¿) E-ß(«o) for t in 5. As in Theorem 1, one can find a constant vector c and a measurable set EES of positive ß measure such that x0±c, for t in E, is contained

License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use 606 SAMUELKARLIN [August

in A. An application of the theorem of Liapounov yields a measur- able subset Eo of E with

Pi(Eo)= I fidp = — I fidp = — pi(E) for i = 1, , n. J B<¡ 2 J B 2 Put c, t in Eo, — c, t in £0, fit) ■c, t in E — Eo, and #(i) = c, tinE— Eo, 0 elsewhere, 0 elsewhere.

It is easy to verify that x0= (£o+

/(*o + $)f>dp= I xofidp+ c\ I /¿¿/i - I /id/t «J LVi:0 J E-E0 J

= I *o/¡¿M+ C — I /.¿M— — I /¿¿Ai

= I *o/i¿M-

Similarly, f(xo-r-$)fidp = fxofidp. Therefore ío+í and *o+^ satisfy the linear inequalities and hence lie in T. We have exhibited a con- tradiction of the extreme point nature of x0 and the proof of the theorem is hereby complete. Although we employed the convexity part of the theorem of Lia- pounov, we can now use Theorem 2 to obtain an extension of the theorem of Liapounov. It is to be remarked, however, that the proof of convexity in the Theorem of Liapounov is the simpler result to obtain and is also used in [l] and [2]. Theorem 3. Let p¡ (j = l, ■■•,«) be atomless, finite measures. Then M[A,pj] = M[B,pj]. Remark 4. An immediate conclusion of this theorem is that the range in Enm of fx(t)dpj(t) where x(t) ranges over Mb is convex and closed. This follows from the evident convexity and closedness of M [A, p¡\ (see Remark 3). Proof. Let x0 be any point in MA- Put \i=Jxodpi. Let T be the set of all * in MA for which li —fxdp{. The set is weak * closed and convex and hence bicompact (see Remark 3). There exists conse- quently an extreme point xi of V. By virtue of Theorem 2, xi is in

License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use 19531 EXTREME POINTS OF VECTORFUNCTIONS 607

Mb and fxidßi=£i=fxodßi. Thus, we have shown that any point in M[A, ßj] is in M[B, ßj\. On the other hand since BEA, evidently we get M[B, ßj]EM[A, ßj]. Combining we have established our re- sult. Corollary 1 [1]. Let C represent any closed in En, then M[C, ßi] is convex and closed if the ßj are atomless. Proof. Let A be the convex span of C. Clearly A is closed and SCC. Evidently, M[B, m]CM[C, in]CM[A, ßi}. However, Theo- rem 3 implies the equality of the outside two sets which thus gives the result (sea Remark 4). As was pointed out in [l] and [2] the closure of M[C, ßi] can be established under any circumstances regardless of the nature of the measures ßi. In fact, any measure ß can be expressed as a sum of an atomless measure ß* and a countable union of pure atomic measures which we designate by ¡i. In notation, ß—ß*-\-ß. In the case of ß one can show directly the closure [l ; 2]. The case of ß* was handled in Theorem 3. Thus, to obtain the closure property for the general case one needs only to apply this decomposition result to the meas- ure ß=ßx+ßi+ ■ ■ ■ +ßn, and invoke the above remarks to the parts of ß. Precisely: M[C, ß*+ß] = convex span (Jl7[c, ß*], M[c, ß]). Closure properties of extreme points. This section investigates whether the set of extreme points of Ma is weak * closed or weakly closed. The measure ß is taken to be atomless and A is assumed to contain more than one point. Theorem 4. Let B = B, then the set of extreme points of MA is not weak * closed and the extreme points Mb and Ma are sequentially weakly closed. Proof. We suppose that the set of extreme points of M a is weak * closed. Let x denote any element of Ma and consider any integrable fxit), • • • ./n(í)- Theorem 2 provides an extreme point x0 such that Jfixdß=ffiXodß for i=x, • • ■ , ». Of course, xo depends on the choice of fi, ■ ■ ■ ,/„. Put Gifa) = [x0\ x0 an extreme point of MA and fxofadß=fxfadß]. The assumption implies that G(/„) is weak * closed and nonempty. The above argument shows that every finite inter- section of G(Jai) is weak * closed and nonempty. Since the extreme points of Ma are weak * closed and hence bicompact, we have L = ClGiJa)5^0. Let xobe in L, then fxfdß=fx0fdp for every integrable / and hence x = x0 almost everywhere. Hence every x in Ma is a point in Mb, which is clearly impossible since A consists of more than one point. We now verify the second assertion of the theorem. If xn con-

License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use 608 SAMUEL KARLIN [August

verges weakly to x, then at least xn converges almost everywhere to x and ||x„|| M(p, %)°°) is defined as the countable infinite direct product of M(p, %). An element of M°° has the form x = (xi(t), x2(t), • ■ ■ ) where each xt(t) is in M(p, %). Let (m) denote the consisting of all bounded sequences. It is well known that (m) is the conjugate space of (/) (sequences which converge absolutely). Therefore bounded weak * closed sets in (m) are bicompact and convex bounded weak * closed sets are spanned in the weak * topology by the extreme points. This is a restatement of the Krein-Milman Theorem. We discuss an example to indicate the extensions of the preceding theory to the infinite case. A. Let x = (xi(t)), i = l, 2, • • • , with 0£xt(ß)£l. Let MA denote the set of all such points in MK. It can be shown in a manner similar to the proof of Theorem 1 that the extreme points of Ma consist of those elements x whose value for almost every t lie in the extreme points of the set B in (m) which consists of sequences (mi) with w, = 0 or 1. Also Ma is weak * closed. Furthermore, it follows that if additional linear conditions jxdp¡ =a¡, j=í, • • • , m, are imposed with pj atomless, then no new extreme points are added. This yields as in Theorem 3 the convexity and weak * closure of the span in (m) of (jxdpj). More detailed results on the infinite vector functions will appear in a later publication.

License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use 19531 EXTREME POINTS OF VECTOR FUNCTIONS 609

Extreme points in measure spaces. It is of interest to compare the type of results obtained above with the description of the set of ex- treme points in other Banach spaces where constraints are present. Specifically, we study the extreme points of the set of all positive regular measures of total variation one defined on a bi-compact Hausdorff space X subject to linear constraints. Let Ai(x), • • ■ , A„(x) be continuous functions defined on X. We look to characterize the set of extreme points of the set L of all posi- tive measures = distributions ß with fdß=i, /A,(x)d/x = a,-. If we con- sider the image M in Euclidean »-space of /¿—>{/A,(x)¿/i}, then it follows easily that M consists of the set of all points, in the convex span of the bi-compact image r of X given in parametric form by i, = A<(x), t = l, • • • , », with x in X. Indeed, let C represent the convex span of Y. As the convex span of the compact set Y, C is closed. If I denotes a point in M but outside C, then there exists a hyperplane which separates C and t. Therefore constants £»,•exist so that E^»'«<— ^<0 ar,d E^<(*)^0 f°r each * m X- Since /< = fhiix)dßix) for some/i, then 0> EW*"" J 2^,bihiix)dßix)^0, which is impossible. This shows that C = M. We now specialize the functions At(x) so that for any choice of « points Xi, x2, • • • , x„ the rank of the matrix A<(x;) is ». Under this condition we now show that points on the boundary of M come from unique measures possessing only a finite number of points of increase. To this end, let I be a point on the boundary of M whose components are given by ti=fhiix)dßoix). There exists a supporting plane to M at /. Hence, f2~^bihiix)dßoix) =0 for appropriate constants and, since E¿>¿A,-(x)^0, we get for every point x of increase of ßo that E^'M*) = 0. This implies in view of the property of A¿(x) that at most » points of increase are possessed by ßo. Since E"-i XiAj(x,) =£/ where hiixj) has rank « we obtain that the X,- are unique. This establishes the assertion that boundary points of M correspond to unique measures. Finally, we investigate the form of the set of extreme points of L. It is clear that the linear constraints /A¿(x)áju(x) =a,- define geo- metrically a section K of the convex set M and that extreme points of K must be boundary points of M. Thus any extreme point of K is built up of a convex finite sum of pure atomic measures. Without loss of generality, we can therefore restrict our considerations only to measures ß which are finite convex combinations of pure atomic measures. Let Ao(x) = l. We now suppose, in addition to the condi- tions stated before, that the rank of the determinant A,(x;) for « + 1 distinct points x¡ is » + 1, i = 0, !,-••,». Suppose now that ß

License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use 610 SAMUEL KARLIN

= Za m L(ai ^0). Clearly any point of increase for a sin- gle pi is shared by p. Let xi, ■ ■ • , xm denote the totality of points of increase of p. Since fhi(x)dp(x) =a¿ for i= 1, • • • , n and Jh0(x)dp(x) = 1, we obtain that

m (*) Z MX*<)= ai> i = o, l, • • •, w, .-i where a0=l. Since the rank of [A/*.-)] is min(« + l, m) it follows that the dimensionality of solutions in X of the linear equations (*) will be m —min (w + 1, m) = max (m — (n + 1), 0). Thus if m^«+l, then the solution is unique and p is an extreme point. If m > w + 1, then since a solution with X,>0 exists, it follows if Z?-i z,Aj(x,) = 0 that X, + ez,- ^0 and ZQ^i±€Zi)hj(Xi) =a¡ for e sufficiently small and hence X< = (Xi+€z¿)/2 + (Xj —ez,)/2. This shows that the extreme points are given by precisely those measures of L which are composed of not more than w + 1 pure atomic measures. We have thus established: Theorem 5. The extreme points of the set L of all positive regular measures over a topological space X satisfying Jdp(x) = 1, fhi(x)dp = ai, i=i, • • • , n, with hi(x) continuous, consist of those measures in L with not more than w + 1 points of increase provided that the rank of {hi(xj)} is w + 1 where i, j = 0, 1, • • • , n, ho(x) = l, and x¡ are distinct points of X. References 1. D. Blackwell, The range of a vector measure, Proceedings of the American Mathe- matical Society vol. 2 (1951). 2. A. Dvoretsky, A. Wald, and J. Wolfowitz, Relations among certain ranges of vector measures, Pacific Journal of (1951). 3. L. Alaoglu, Weak topologies of normed linear spaces, Ann. of Math. (1940). 4. A. Liapounov, Sur les fondions-vecteurs complement additives, Bull. Acad. Sei. USSR. vol. 4 (1940). California Institute of Technology

License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use